Estimation of soil temperature using gene expression programming and artificial neural networks in a semiarid region

Soil temperature (T _s) is one of the most important parameters which affect physical and chemical properties of soil. In the present study, two biologically inspired approaches for artificial intelligence including gene expression programming (GEP) and artificial neural networks (ANN), as well as multiple linear regression (MLR) were used to estimate the soil temperature at six different depths (5, 10, 20, 30, 50 and 100 cm) for the Sanandaj synoptic station in a semiarid region in western Iran. Twelve combinations of meteorological parameters, such as minimum and maximum air temperatures, relative humidity, wind speed, sunshine hours and extraterrestrial radiation, were used as input variables. The full data set containing soil temperature and atmospheric parameters, which spans the time period from 1997 to 2008, was divided into training (1997–2004) and testing (2005–2008) data sets. To evaluate the accuracy of the models, determination coefficient (R ²) and root mean square error (RMSE) were calculated. The results showed that the GEP, ANN and MLR were able to model T _s at different depths. However, the performance of the ANN approach was the best.